Python → ML → deep learning

In plain words

Python is a programming language widely used for data analysis and AI. pandas is an add-on library for Python that handles tables of data, a bit like a spreadsheet you control by writing instructions instead of clicking. Learning to load, clean and summarise a table is the first step in this track. Most of the later steps assume you can do it.

Healthcare example (illustrative)

Imagine a renal registrar with an audit spreadsheet of dialysis sessions exported from the hospital system. With a short pandas script they load the file, remove duplicate rows, flag missing blood results, and count sessions per patient per month. Next month they re-run the same script on the new export and get results made in exactly the same way. Every step is written down, so a colleague can check it.

A little more technical

pandas stores a table as a "DataFrame". The everyday operations are filtering rows, selecting columns, grouping and summarising, joining two tables on a shared identifier, and handling missing values. Much of real analysis time goes on cleaning rather than modelling. Work only with de-identified data in an approved environment. Never paste patient data into a consumer chatbot for coding help (see Track 1 on PHI rules).

Don't learn this yet

This whole track is optional: skip it unless you are building models or doing technical research. If you do start, skip object-oriented programming, web development and speed optimisation. Loading, cleaning and summarising tables is enough for now.

Go deeper (optional)

• Python for Data Science (noc26_cs80) #144

In plain words

scikit-learn is a Python library of standard ("classical") machine-learning methods, such as logistic regression, decision trees, random forests and clustering. It also includes tools to split data and measure performance. This is where you learn the core discipline of ML: train on one set of data and judge the model on data it has never seen.

Healthcare example (illustrative)

Imagine a hypothetical research project in an emergency department that predicts hospital admission from triage data. The team fits a logistic regression and a random forest on past visits, then tests both on a separate set of later visits. They check discrimination and calibration (Track 2). They also spot that one variable, the admitting ward, is only recorded after the decision; including it would be leakage and would make the model look far better than it really is.

A little more technical

The usual pattern is: split the data, fit the model on the training part, predict on the held-out part, and evaluate. Cross-validation repeats this over several splits to give a steadier estimate. "Pipelines" keep data-preparation steps inside the training data, which helps prevent leakage. Any result still needs external and temporal validation before anyone relies on it (Track 3).

Don't learn this yet

Skip this entirely unless you are building or researching. If you are, leave deep learning, automated tuning tricks and exotic algorithms until the basics feel routine.

In plain words

This step is a Stanford course on Coursera that teaches core ML concepts for clinical problems (see #43). The resource list records it as intermediate level, with no coding required, light maths, and classical (pre-LLM) content. It belongs to Stanford's five-course AI in Healthcare specialisation, which is designed for clinicians and computer-science professionals (see #42). On this path it is the conceptual bridge: after hands-on basics, you learn how clinical questions are turned into ML problems.

Healthcare example (illustrative)

Imagine a haematology registrar who wants to predict sepsis in patients with neutropenia. The course helps them frame the problem properly: how exactly the outcome is defined, the time point at which the prediction is made, and which data would really be available at that moment. Getting these right matters more than which algorithm they choose.

A little more technical

It suits clinicians who want a structured, conceptual course, and researchers who will work with data scientists. Reviewers describe the specialisation's videos as lecture-heavy, and its content is classical-era, so pair it with Track 6 for LLMs. The specialisation's capstone is a guided patient journey (a COVID-19 respiratory scenario) over a de-identified EHR and imaging dataset, covering risk stratification and regulatory and ethical limits (see #44). The resource list records CME for these courses; check the details on the current course page.

Don't learn this yet

Track 5 is optional: skip it unless you are building or researching. If your goal is to judge AI tools rather than build them, Track 3 serves you better than this course.

Go deeper (optional)

• Fundamentals of Machine Learning for Healthcare #43
• AI in Healthcare Specialization (5 courses) #42
• AI in Healthcare Capstone #44

In plain words

This is a three-course specialisation from DeepLearning.AI on Coursera (see #46); its courses on diagnosis, prognosis and treatment can also be taken individually (see #47). The resource list records it as intermediate, requiring intermediate Python and statistics and probability, at about two months at 10 hours a week. It is the first step on this path where you train models on medical tasks yourself. The resource list picks it for coding in clinical ML because of its real medical tasks and graded labs.

Healthcare example (illustrative)

Imagine a cardiology fellow with some Python who wants to model time to readmission in heart failure. The course's survival-modelling material shows why an ordinary yes/no classifier handles patients who leave follow-up early badly, and how survival models deal with them. The fellow would still validate any model carefully before suggesting it influence care.

A little more technical

The topics listed are chest X-ray classification, 3D MRI tumour segmentation (U-Net), survival models (Cox and random survival forests), treatment-effect estimation from randomised trials, model interpretation, and NLP labelling of radiology reports. The content is classical and pre-LLM, with no generative AI. On this path it comes before MIT 6.7930: the applied course first, then the rigorous one.

Don't learn this yet

Skip this unless you are building or researching. If you only want to understand imaging or prognostic AI well enough to appraise it, Tracks 2 and 3 are the better use of time.

Go deeper (optional)

• AI for Medicine Specialization (3 courses) #46
• AI for Medical Diagnosis / Prognosis / Treatment #47

In plain words

This is MIT's graduate-level Machine Learning for Healthcare course (6.7930/HST.956), with free course materials (see #48). The resource list records it as advanced, requiring both coding and maths. It covers clinical data, risk stratification, disease progression, precision medicine, causality, interpretability, fairness and time series. On this path it is where you move from using methods to understanding their limits on clinical data.

Healthcare example (illustrative)

Imagine an intensive care researcher who wants to know from routine records whether earlier referral to critical care outreach improves outcomes. Sicker patients tend to be referred earlier, so a simple comparison could make early referral look harmful. The course's material on causality helps them recognise this confounding and choose a more defensible analysis.

A little more technical

It suits researchers and builders who are already comfortable with Python and statistics. An older 2019 version on MIT OpenCourseWare has full notes and lectures but is pre-LLM (see #49). A separate Fall 2024 MIT course covers the practical side of putting ML into healthcare use: robust, private and fair ML on real retrospective data, visualisation, and implementation science (see #50). The roadmap already points to this course's first lecture in Track 2.

Don't learn this yet

Skip this track unless you are building or researching. Even then, don't start here without Python, scikit-learn, the Track 2 statistics and the applied DeepLearning.AI course behind you.

Go deeper (optional)

• 6.7930/HST.956 Machine Learning for Healthcare #48
• 6.S897/HST.956 Machine Learning for Healthcare (Spring 2019) #49
• HST.953 Clinical Data Learning, Visualization, and Deployments (Fall 2024, MIT OpenCourseWare) #50

In plain words

CNNs (convolutional neural networks) are neural networks designed for images. They are used for classification (is there an abnormality?), detection (where is it?) and segmentation (outlining a structure pixel by pixel). DICOM is the standard format for medical images and their accompanying information, and PACS is the hospital system that stores and displays them. MONAI is an open-source toolkit for medical-imaging AI, with free tutorials on 2D and 3D segmentation and classification, most of which run in the browser through Colab (see #54).

Healthcare example (illustrative)

Imagine a radiotherapy team that wants to test automatic outlining of organs at risk on planning CT scans. With approval, it exports de-identified DICOM images from PACS and trains a segmentation model using MONAI. In use, the model proposes outlines, and the clinical oncologist reviews and edits every one before the plan is approved.

A little more technical

DICOM files carry patient identifiers in their headers, so de-identification is a key step before any research use. Segmentation is usually judged by how much the model's outline overlaps an expert's (for example, the Dice score), but a good average can hide clinically important misses. 3D scans need far more computing power than 2D images. For broader foundations, the resource list includes courses on deep learning for computer vision, which assumes prior ML (see #148), and on medical image analysis (see #139).

Don't learn this yet

Skip this unless you are building imaging AI or doing imaging research. Even then, leave PACS integration, DICOM networking, speed-up techniques and newer imaging architectures until you have trained and evaluated a simple model end to end.

Go deeper (optional)

• MONAI tutorials #54
• Deep Learning for Computer Vision (noc25_cs93) #148
• Medical Image Analysis (noc25_bt61) #139
• AI for Medicine Specialization (3 courses) #46